120 likes | 252 Vues
This study focuses on enhancing processing techniques for the effective classification of instruments in polyphonic music. Given the challenges of overlapping tones and octave confusion, traditional methods fall short. By integrating a Missing Feature Approach and spectral analysis, the project aims to improve accuracy and reliability in identifying instruments across various families. Initial testing will occur in monophonic music settings, followed by analysis in polyphonic contexts. The ultimate goal is to refine automatic music transcription processes through robust instrument classification.
E N D
Instrument Classification in a Polyphonic Music Environment Yingkit Chow Spring 2005
Objective • To develop and expand on processing techniques for classification of instruments in a polyphonic music environment Incentive • Classification of instruments can prove useful for automatic music transcription.
General Background • Most previous research with instrument classification works with monophonic audio • Work in Polyphonic music usually deals with “Blackboard” system that accesses knowledge sources and works with a top-down approach • Difficulties with polyphonic music: • Overlap of tones can be detrimental to extracting the correct identifying features • Confusion of octave
Features Used for Classification in Monophonic Music Analysis [1] Features Used: • RMS Envelope • 60 % Accuracy, 60% Reliability • CQT Frequency Spectrum (PCA) • 66 % Accuracy, 68% Reliability • MSA Trajectories (PCA) • 75 % Accuracy, 76 % Reliability • Combined: • 82% Accuracy, 83% Reliability
Results from [1], Monophonic Audio • CONFUSION MATRIX NNC Combined: k=5 WEIGHTED MAJORITY, Confusion Matrix weighted (reliability)
Missing Feature Approach[2],For Polyphonic musicJana Eggink and Guy J. Brown • Find fundamental frequency and compare against harmonic sieves • Spectral peaks of fundamental and harmonics are the features used for classification (50-6kHz, 60 Hz window) • Instruments tested in this paper: • (flute, clarinet, violin, cello, oboe)
Missing-Feature Approach, Special Conditions[2] • Frequency regions with energy from interfering tones are excluded from classification process. • Cepstral coefficients are not used as features since they correspond to all frequencies. Local spectral features are used. • Cannot handle drum or “untuned percussion instrument”
Results with Missing Feature Approach to monophonic music • Confusion Matrix for the 5 instruments in a Monophonic music environment • The identification of the family of instrument (String, Woodwind) is about 85%
Missing Feature Approach,Instrument Classification in 2 tone samples • Confusion Matrix for the 5 instruments in a polyphonic music environment (2 simultaneous tones)
Project Goal • I will test the Missing Feature Model against a different set of instruments: • Instruments to be selected based on available sample set and to provide a variety of instruments from different families. • Alternative source would be taking input from a synthesized version of the instrument (MIDI to Wav). • Add system over the missing feature model to include information from neighboring frames (in time) and use information of partials in the classification scheme
Testing • I will test my system first, within monophonic music, to get an upper bound on the capabilities of the system for each instrument. • Secondly, I will test the system within a 2 note environment
References 1. “Multi-feature Musical Instrument Sound Classifier”, I. Kaminskyj, http://www.mikropol.net/volume6/kaminskyj_i/kaminskyj_i.html 2. “A Missing Feature Approach to Instrument Identification in Polyphonic Music”, by Jana Eggink and Guy J. Brown http://www.dcs.shef.ac.uk/%7Ejana/egginkICASSP03.pdf 3. “Instrument Recognition in Acompanied Sonatas and Concertos”, by Jana Eggink and Guy J. Brown http://www.dcs.shef.ac.uk/%7Ejana/egginkICASSP04.pdf 4. Music Samples from the University of Iowa, http//:theremin.music.uiowa.edu/